Grand Challenges in Agentic AI for Cyber Operations: A Research Agenda

Abstract

Agentic AI systems are reshaping cyber operations at a pace that outstrips the mechanisms needed to deploy them responsibly. The first documented autonomous cyber attack, in September 2025, demonstrated that the technology has crossed the threshold from research capability to operational threat, yet the technical robustness, human-AI trust, and governance frameworks required for responsible adoption remain underdeveloped. This paper argues that the resulting gaps constitute a control deficit manifested through a set of grand challenges across four reinforcing dimensions: technical limitations and vulnerabilities, the trust deficit between operators and AI agents, insufficient governance, and dual-use escalation risks. Through a structured expert analysis that draws on operational, technical, human, and policy perspectives, we characterize the interactions among these dimensions. We then propose a research agenda to help the cyber operations community address these grand challenges in a coordinated manner, accounting for both the dependencies among dimensions and the operational urgency posed by the adversary's adoption of the same technology. In doing so, the paper calls on researchers, practitioners, and policy-makers to collectively shape the responsible integration of agentic AI into cyber operations without ceding the advantages it provides.